[LLVM] Bump LLVM to 20.1.8

[gbaraldi]

Notably this has the IRCE fix

[giordano]

Notably this has the IRCE fix

I presume this refers to #48918? Is that issue fully fixed or something like #59129 is also required? Any regression tests we can add here?

[giordano]

Analyzegc is unhappy https://buildkite.com/julialang/julia-master/builds/49705#01988187-7f07-4890-b9af-a7b09b133eac/1171-1767

    ANALYZE src/clang-sagc-gc-stock
/cache/build/tester-amdci5-12/julialang/julia-master/src/gc-stock.c:3080:9: error: Calling potential safepoint as SimpleFunctionCall from function annotated JL_NOTSAFEPOINT [julia.GCChecker]
 3080 |         gc_mark_loop(ptls);
      |         ^~~~~~~~~~~~~~~~~~

[Keno]

Is that issue fully fixed or something like #59129 is also required?

Would also fix it, but this fixes the observed crash, but in the case where the crash happens, the original semantics will cause the value to be rooted longer than intended.

[KristofferC]

Would it be good to have an in repo test of the issue that this bump fixed? So we notice if it regresses.

[gbaraldi]

We never got a really small reproducer sadly. It needs at least a big chunk of staticarrays

[Keno]

Yeah, it's somewhat hard to come up with a small reproducer. It needs at least control flow outside a @GC.preserve region (or something that LLVM will turn into that) and an IRCE-eligible loop. It should be possible to write such a reproducer now that we understand the issue, but I don't have one.

[gbaraldi]

We can write an LLVM test and run it locally but keno's patch tests this upstream